Build tensorflow from source

Hello, I am trying to build tensorflow from source on a small Rocks cluster. Unfortunately, when I use my own tensorflow build I get only 30 % GPU utilization. For testing purposes I use cifar10 dataset, which was cloned using git clone command. On the other hand, when I am trying to execute the same script inside a tensorflow docker container I get over 60 % of GPU utilization. May someone give me some hints how can I boost GPU utilization or where does docker saves information about how docker container was built (in my case